Joint genotypic and phenotypic outcome modeling improves base editing variant effect quantification

Abstract

CRISPR base editing screens enable analysis of disease-associated variants at scale; however, variable efficiency and precision confounds the assessment of variant-induced phenotypes. Here, we provide an integrated experimental and computational pipeline that improves estimation of variant effects in base editing screens. We use a reporter construct to measure guide RNA (gRNA) editing outcomes alongside their phenotypic consequences and introduce base editor screen analysis with activity normalization (BEAN), a Bayesian network that uses per-guide editing outcomes provided by the reporter and target site chromatin accessibility to estimate variant impacts. BEAN outperforms existing tools in variant effect quantification. We use BEAN to pinpoint common regulatory variants that alter low-density lipoprotein (LDL) uptake, implicating previously unreported genes. Additionally, through saturation base editing of LDLR, we accurately quantify missense variant pathogenicity that is consistent with measurements in UK Biobank patients and identify underlying structural mechanisms. This work provides a widely applicable approach to improve the power of base editing screens for disease-associated variant characterization.

Genetic variation contributes substantially to disease risk. While well-powered genome-wide association studies (GWAS)¹ and rare variant analyses from cohort studies such as the UK Biobank (UKB)² have associated thousands of loci and genes with clinical phenotypes, these observational approaches are often insufficient to pinpoint causal variants. Multiplexed assays of variant effect³ such as deep mutational scanning⁴, saturation mutagenesis⁵, massively parallel reporter assays⁶ and CRISPR-based screens^7–9 enable high-throughput evaluation of the causal impact of individual variants. CRISPR base editors, fusions of Cas9 nickase and single-stranded cytosine or adenine deaminase enzymes^10,11, enable site-specific installation of transition variants (A>G, C>T), which comprise the majority of disease-associated variants, in their endogenous genomic context¹². Base editing screens have been employed to dissect coding and noncoding variant effects^13–28. However, base editing efficiency and allelic outcomes vary substantially depending on the local sequence context surrounding the target base, the editor used and the cellular context²⁹. This variability has confounded the analysis of base editing screens. Prior efforts have enabled accurate prediction of editing outcomes based on reporter assays²⁹, but these predictions do not generalize well to unprofiled base editors and cellular contexts²⁰.

Two recent studies have begun to address this shortcoming by simultaneously profiling phenotypic impacts of each gRNA along with editing outcomes using a base editor reporter (or sensor) assay^20,21. However, these studies only use editing outcomes to filter out gRNA species with low editing efficiency and do not account for the remaining genotypic heterogeneity when inferring variant-induced phenotypic effects.

Here, we design an experimental–computational pipeline to improve the accuracy of variant effect estimation in base editing screens. By incorporating a target site reporter sequence into the gRNA construct, we simultaneously measure the editing outcomes and phenotypic impacts of each gRNA. We develop a computational pipeline, BEAN, that normalizes and deconvolves the phenotypic scores of target variants by using genotypic outcomes collected from the target site reporter and by sharing information among neighboring gRNA species. BEAN provides an integrated solution to experimental assessment of variant effects through base editing screens. We systematically benchmark BEAN against current state-of-the-art methods for the analyses of pooled CRISPR screens and show substantially improved performance of BEAN.

To leverage activity-normalized base editing screening, we have conducted screens assessing the impact of LDL cholesterol (LDL-C)-associated GWAS variants and LDL receptor (LDLR) coding variants on LDL-C uptake in HepG2 hepatocellular carcinoma cells. The serum LDL-C level is both a clinically important factor that contributes to coronary artery disease (CAD) risk and a commonly cataloged phenotype in biobanks with high-quality human quantitative trait data. A trans-ancestry GWAS meta-analysis has identified >400 loci associated with LDL-C³⁰. Yet, the causal variants and mechanisms by which many of these loci modulate LDL-C levels remain unknown.

LDL-C levels are also impacted by rare coding variants. In severe instances, inherited monogenic variants in several genes cause familial hypercholesterolemia (FH), a disease associated with extremely elevated LDL-C levels and premature cardiovascular disease³¹. The majority of genetic mutations known to cause FH occur in LDLR, encoding a cell surface receptor that takes up LDL, thus removing it from circulation³². Despite the effectiveness of lipid-lowering therapies, patients with FH are still two- to fourfold more likely to have coronary events than the general population³³. Elevated LDL-C levels increase cardiovascular disease risk throughout life; therefore, the early identification of at-risk individuals would have immense clinical utility³¹. However, many LDLR variants currently lack clinical interpretation. Of the 1,427 LDLR missense variants in the ClinVar data-base³⁴, 50% are classified as variants of uncertain significance (VUS) or have conflicting interpretations of pathogenicity (‘conflicting’), thus impeding the use of genetic information in the diagnosis of hypercholesterolemia. Likewise, of the 758 unique LDLR missense variants carried by sequenced individuals in the UKB cohort, 69% are either unreported or have an uncertain annotation in ClinVar. Altogether, improved understanding of LDLR variant impacts would enable earlier diagnosis and treatment for a large number of individuals at risk for hypercholesterolemia and FH.

We sought to understand the causal variants underlying LDL-C levels and their mechanisms by investigating both common GWAS-associated variants and rare LDLR coding variants. To this end, we performed a pooled base editing screen in HepG2 cells in which cells were flow sorted by the uptake of fluorescent LDL-C levels. Our system provides a scalable assay to assess variant impacts on serum LDL-C levels³⁵, given that the majority of serum LDL-C is cleared in the liver³⁶. By applying our experimental computational pipeline to these screens, we identify LDL uptake-altering GWAS-associated variants and characterize their downstream impacts to nominate causal variants that act through genes not previously implicated in LDL-C control. Through saturation tiled base editing of LDLR, we find strong correlation between inferred functionalscores and UKB patient measurements and reveal a key, conserved tyrosine residue in each LDLR class B repeat that interacts with the neighboring repeat to maintain protein structural integrity. Altogether, BEAN provides a widely applicable tool to characterize single-nucleotide variant function.

Results

A base editing reporter profiles endogenous editing outcomes

To enable accurate interrogation of variant effects at scale, we built a platform to perform dense, high-coverage base editing screens that accounts for variable editing efficiency and genotypic outcomes. To maximize coverage of variants in base editing screens, we built lentiviral adenine (ABE8e)^11,37 and cytosine (AID-BE5)²⁹ deaminase base editor constructs using the near-PAM (protospacer adjacent motif)-less SpCas9 variant SpRY³⁸. Both base editors showed native genomic editing activity, with ABE8e–SpRY showing considerably more robust maximal activity (Supplementary Fig. 1a). Editing efficiency was increased by 5–10% by prior lentiviral integration of constitutively expressed base editors and by transient dosing of cells with the histone deacetylase valproic acid³⁹ immediately after base editor and gRNA transduction (Supplementary Fig. 1b,c).

To account for variable base editing efficiency^29,40,41, we synthesized and cloned each gRNA paired with a 32-nucleotide reporter sequence comprising the genomic target sequence of that gRNA into lentiviral base editor vectors (Fig. 1a and the Methods), akin to previously published base editing outcome reporter (‘sensor’) constructs^20,21,29. When introduced into cells, the gRNA can edit both its native genomic target site and the adjacent target site (reporter) in the lentiviral vector, which can be read out using next-generation sequencing (NGS).

Fig. 1 |. Activity-normalized base editing screening pipeline.

a, Schematic of the activity-normalized base editing screening process and analysis by BEAN. A library of gRNA species, each paired with a reporter sequence encompassing its genomic target sequence, is cloned into a lentiviral base editor expression vector. Lentiviral transduction is performed in HepG2 cells, followed by sorting four populations by flow cytometry based on fluorescent LDL-C (BODIPY-LDL) uptake. The gRNA and reporter sequences are read out by paired-end NGS to obtain gRNA counts and reporter editing outcomes in each flow cytometric bin. BEAN models the reporter editing frequency and allelic outcomes and gRNA enrichments among flow cytometric bins using BEAN to estimate variant phenotypic effect sizes. b, Schematic of the LDL-C variant library gRNA design for selected GWAS candidate variants with a Manhattan plot showing variant P values from a recent GWAS study⁸⁶ (P values are for the logistic regression coefficient being nonzero; see ref. 86 for details). gRNA species tile the variant at five positions with maximal editing efficiency (protospacer positions 4–8). M, million; chr, chromosome. c, gRNA coverage of the LDLR tiling library across the LDLR coding sequence along with 5′ and 3′ untranslated regions and several regulatory regions. Coordinates shown are of hg19. The left diagram was created with https://www.biorender.com. d, Average editing efficiency of ABE8e–SpRY by protospacer position and PAM sequence. e, Adjacent nucleotide specificity of ABE8e–SpRY editing represented as a sequence logo from 7,320 gRNA species; the height of each base represents the relative frequency of observing each base given an edit at position 0. f, Scatterplots of per-nucleotide editing efficiencies in the reporter and endogenous target sites. All edits introduced by each of 49 gRNA species across four loci and three experimental replicates are plotted separately. The number of nucleotide edits across the three replicates is reported in each panel as n. The accessibility of the four loci as measured by ATAC-seq (assay for transposase-accessible chromatin with sequencing) signal in HepG2 cells is shown at the bottom, and scatterplot markers are colored by the accessibility of each nucleotide. Pearson correlation coefficients are shown as r.

We designed two gRNA libraries using this approach to improve understanding of the genetics of LDL-C levels. The first library (LDL-C GWAS library) targets 583 variants associated with LDL-C levels from the UKB GWAS cohort² (Supplementary Note 1 and Supplementary Table 1), also including positive control gRNA species that ablate splice donor and acceptor consensus sites in six genes (LDLR, MYLIP, ACAT2, SREBF2, HNF4A, LSS) found to have significantly altered LDL-C uptake upon knockout³⁵ and 100 nontargeting negative control gRNA species, totaling 3,455 gRNA species. We designed five tiled gRNA species for each variant allele that place the variant in positions shown to induce most efficient editing with ABE8e⁴² (Fig. 1b).

The second library (LDLR tiling library) targeted the LDLR gene (Supplementary Table 2) with every possible gRNA targeting the LDLR coding sequence on both strands as well as lower-density gRNA species targeting the 50 nucleotides flanking each LDLR exon, LDLR 5′ and 3′ untranslated regions, the promoter and two intronic enhancers (Fig. 1c) as well as 150 nontargeting negative control gRNA species, resulting in a total of 7,500 gRNA species.

We first profiled editing outcomes of ABE8e–SpRY and AID-BE5–SpRY in HepG2 cells through NGS of >10,000 gRNA–reporter pairs consisting of two gRNA libraries designed to target LDL-C altering candidate variants through an end-to-end computational toolkit for base editing screens, BEAN (Supplementary Note 2). The result clearly recapitulated the hallmark positional preferences of these base editors^29,37, the NRY (N: A, T, C, G; R: Purines, Y: Pyrimidines) PAM preference of the SpRY enzyme³⁸ and the relative depletion of editing at AA dinucleotides by ABE8e²⁹ (Fig. 1d,e and Extended Data Fig. 1). Notably, the average maximal positional ABE8e–SpRY editing frequency at protospacer positions 3–8 across dinucleotide PAM sequences ranged from 32% to 46%, indicating the ability of this enzyme to install variants efficiently across a wide variety of genomic locations.

To validate that editing of the reporter provides an accurate surrogate for endogenous editing, we sequenced both the reporter and endogenous target sites for 49 gRNA species across four loci surrounding LDLR with varying levels of HepG2 chromatin accessibility (Supplementary Table 3). We demonstrate that nucleotide-level and allele-level reporter editing fractions correlate well with endogenous target site editing fractions (Fig. 1f and Extended Data Fig. 2; average Pearson correlation across four loci for per-nucleotide editing rate r = 0.70, per-allele editing rate r = 0.69), and the reporter showed higher correspondence than BE-Hive predictions²⁹ (nucleotide r = 0.44, allele r = 0.64) (Supplementary Figs. 2 and 3). Notably, while reporter editing correlated with endogenous editing at all four loci, we found that endogenous editing frequency also depended on the accessibility of the target region (Fig. 1f), as has been previously reported for Cas9 nuclease^43–45 and base editors^40,41.

Having profiled the relationship between endogenous and reporter editing, we performed fluorescent LDL uptake screens with each library with the aim of using genotypic outcome information from the reporter to improve screen analysis. We hypothesized that using a greater number of sorting bins during flow cytometric isolation would enable the identification of variants with weak effect size resulting from base editing, in contrast to those induced by Cas9 or transcriptional activation or deactivation by CRISPR activation^46,47 or inactivation⁴⁸ (CRISPRa or CRISPRi). Our computational simulations (Supplementary Note 3) supported the hypothesis, showing that a sorting scheme with four bins (very low (0–20th percentile), low (20–40th percentile), high (60–80th percentile) and very high (80–100th percentile) for LDL uptake) was more effective than the commonly used two-bin scheme of low (for example, 0–30th percentile) and high (for example, 70–100th percentile) (Supplementary Fig. 4). Consequently, we performed the screens through flow cytometric isolation of four populations per replicate with the aforementioned four-bin sorting scheme. We observed robust replicability (median Spearman ρ = 0.84 for the LDL-C GWAS library, ρ = 0.88 for the LDLR tiling library) in gRNA counts across replicates (Supplementary Fig. 5), indicating technical reproducibility.

Activity-normalized base editing screen analysis with BEAN

We developed a new Bayesian analysis method, BEAN, to quantify the phenotypic effect of each variant from gRNA abundance in sorted populations using the genotypic outcome information provided by reporter editing combined with the chromatin accessibility data from target cells. Instead of directly using reporter genotype to quantify phenotype, which could lead to underestimation of variant effects (Supplementary Note 4), BEAN assumes that the observed phenotypic distribution in a population of cells for each gRNA derives from a mixture of cells with unedited and edited alleles (Fig. 2). Because multiple gRNA species may induce the same genotypic outcome at different frequencies, BEAN uses this redundancy to build confidence in the predicted phenotypic impacts of a given genotype. As the output for each variant, BEAN provides its effect size that is the posterior mean phenotypic shift along with the corresponding z score and the 95% credible interval (CI). We also note that BEAN can be adapted to alternative experimental designs (Supplementary Note 2).

Fig. 2 |. BEAN models variant effects from activity-normalized base editing screens.

Simplified schematic of the BEAN Bayesian network that models input reporter editing outcomes and gRNA counts. The Bayesian network model recapitulates the data generation process starting from a variant-level phenotype Y_v and models per-gRNA phenotypes (Y_g) as a Gaussian mixture distribution of edited and unedited (wild-type) allele phenotypes. The weights of the mixture components are modeled to generate reporter editing outcomes. gRNA abundance in each sorting bin is then calculated by discretizing the gRNA phenotype based on the experimental design into phenotypic quantiles and is modeled to generate the observed gRNA counts using an overdispersed multivariate count distribution (Methods). BEAN outputs the parameters of the posterior distribution of the mean phenotypic shift as a Gaussian distribution with mean ${\widehat{\mu }}_{\mu }$ (effect size), along with negative control adjusted z scores and CIs, where $\mathcal{D}$ is the input data. WT, wild type.

BEAN identifies LDL uptake-altering GWAS variants

We applied BEAN to the LDL-C GWAS library screen. While variant editing efficiency per gRNA is highly variable with an average edit fraction of 34.0%, most target variants are edited at high efficiency by at least one of the five targeting gRNA species (median maximal editing of 60.4%; Supplementary Fig. 6).

First, we compared the performance of BEAN and five published CRISPR screen analysis methods in distinguishing the effects of positive control splice-altering variants versus negative control nontargeting gRNA species^23,49–52 (Supplementary Note 5 and Fig. 3a). To dissect the contributions of individual features to BEAN performance, we included two reduced versions of BEAN: one that considers reporter editing but not chromatin accessibility (BEAN-Reporter) and another that ignores the reporter, assuming uniform gRNA editing efficiency (BEAN-Uniform) (Supplementary Note 2 and Extended Data Fig. 3). BEAN outperforms other evaluated methods in classifying positive control splicing variants of the strong LDL-C uptake regulators encoded by LDLR and MYLIP (Fig. 3b) as well as those of four additional LDL-C uptake modulating genes³⁵ (Extended Data Fig. 4) against the negative control variants. This improved performance was accentuated when the data were subsampled for fewer replicates, demonstrating BEAN’s ability to maintain robustness even with less data. Importantly, BEAN showed improved performance (mean area under the precision–recall curve (AUPRC) = 0.90 across 15 two-replicate subsamples) over BEAN-Reporter (mean AUPRC = 0.87), which in turn outperformed BEAN-Uniform (mean AUPRC = 0.85), supporting the value of accurately modeling target site editing. Intriguingly, even BEAN-Uniform outperformed alternative approaches, likely due to more accurate modeling of sorting bins, suggesting the utility of BEAN in sorting screens without a reporter.

Fig. 3 |. BEAN improves variant impact estimation from the LDL-C GWAS library screen.

a, Ridge plot of BEAN z-score distributions of positive controls, negative controls and test variants. b, AUPRC for classifying LDLR and MYLIP splicing variants versus negative controls. Metrics for all five replicates are shown as crosses, and metrics of 15 two-replicate subsamples among the five replicates are shown as box plots. Metrics are shown for BEAN, MAGeCK-RRA⁴⁹, MAGeCK-MLE⁵⁰, CRISPRBetaBinomial (CB2)⁵¹, efficiency-corrected log fold change (EC-LFC)²³ and log fold change (LFC). c, Scatterplot of variant effect size and significance estimated by BEAN. Labels show rsIDs of selected variants and dbSNP gene annotations and a manual annotation for the APOB enhancer in parentheses. d, Spearman correlation coefficient between BEAN effect size and log fold change of LDL-C uptake following individual testing of 26 gRNA species. Metrics for all five replicates are shown as crosses, and metrics of 15 two-replicate subsamples among the five replicates are shown as box plots. e, Scatterplot of BEAN effect size and log fold change of LDL-C uptake following individual testing of 26 gRNA species. The Spearman correlation coefficient is denoted as ρ. For statistics in box plots, please see Statistical note for plots.

BEAN identified 54 variants that significantly alter LDL-C uptake (95% CI does not contain 0; Supplementary Table 6). These variants include intronic variants in well-known LDL-C uptake mediators³⁵ such as those encoded by ABCA1, LDLR and SCARB1 (Fig. 3c) and coding or intronic variants in APOE, CCND2, GAS6 and FBLN1 with strong genetic likelihood of causality (UKB SuSiE fine-mapping⁵³ PIP (posterior inclusion probability) > 0.99 and/or the only variant in a fine-mapped credible set³⁰), indicating that the effect of these variants on serum LDL-C was at least partially mediated by LDL-C uptake.

To validate the inferred effect sizes, we performed individual transduction of gRNA species targeting 22 variants and four positive controls with six biological replicates (Methods and Supplementary Table 7). Individual LDL-C uptake log fold change values showed strong correlation with the BEAN effect sizes (Fig. 3d,e, Extended Data Fig. 5 and the Methods), showing that BEAN enables accurate inference of variant effects on LDL-C uptake over a wide dynamic range.

To gain insight into a set of 20 variants for which the mechanism of LDL-C uptake alteration is less clear, we developed a pipeline to assess the cellular mechanism of edited variants (Fig. 4a). First, we asked which of these variants impact chromatin accessibility through pooled variant editing followed by allelic representation analysis comparing ATAC-seq and genomic DNA (gDNA) sequencing. (Fig. 4b and the Methods). A similar procedure has been performed to assess differential allelic chromatin accessibility⁵⁴ and transcription factor binding⁵⁵ but not after base editing.

Fig. 4 |. Functional characterization of LDL-C GWAS variants.

a, Schematic of potential variant mechanisms and figure panels showing data from each mechanistic experiment. b, Schematic of allelic representation analysis to identify variants impacting accessibility. Differential representation of alleles in gDNA and ATAC-seq reflects differential accessibility induced by the base edit or heterozygous reference alleles. RT–qPCR, quantitative reverse transcription PCR. c, Change in ATAC-seq enrichment from the pooled ATAC-seq experiment. Enrichment values are plotted as bars, and 95% confidence intervals are shown as error bars. ‘Edited variant’ denotes enrichment by the base edit, and ‘allele’ denotes enrichment by either of the heterozygous alleles, when available, represented as the enrichment of major (Maj) to minor (Min) alleles. Variants for which the base edit is conducted from the minor allele are shown as red in the color bar. FWER with Benjamini–Hochberg multiple correction is displayed as text for each enrichment value where FWER < 0.1. Three experimental replicates across two conditions were used to obtain the enrichment (Methods). d, Genomic tracks for three selected variants. Multiple transcript variants of RGS19 and OPRL1 are shown in the middle. DNaseIHS, DNase 1 hypersensitivity; GTEx, Genotype–Tissue Expression. e, Fraction of ZNF329 minor allele haplotype reads in gDNA and complementary DNA from untreated HepG2 cells and HepG2 cells with rs35081008^Min>Maj base editing measured using Sanger sequencing. Rep, replicate. f, Change in gene expression following base editing of three selected variants from minor to major alleles. P values of one-sample Student’s t-test of log fold change versus a mean of 0 that are smaller than 0.05 are shown above each bar. Error bars show the 95% confidence interval of bootstrapping. n = 7 experimental replicates for RGS19; n = 8 for the rest. g, Change in cellular LDL-C uptake following CRISPRa or CRISPRi of proximal genes for three selected variants. P values of one-sample Student’s t-test of log fold change versus a mean of 0 that are smaller than 0.05 are shown above each bar. Error bars show the 95% confidence interval of bootstrapping. n = 12 experimental replicates for CRISPRa, n = 6 for CRISPRi. h, Summary schematic of characterization results. For statistics in box plots, please see Statistical note for plots.

We identified five variants as chromatin accessibility quantitative trait loci (caQTL)⁵⁶ among the eight tested variants that are heterozygous in HepG2 cells (Fig. 4c). Two of these loci (rs35081008 and rs2618566) were also identified as caQTL in a recent analysis of 20 human liver tissue samples^30,57. As caQTL analysis cannot isolate the causal effect from linked variants, we evaluated differential accessibility by base editing. Of the 15 variants without technical issues (Methods), eight significantly altered chromatin accessibility when edited (family-wise error rate (FWER) of 0.1). Four variants (rs11149612, rs35081008, rs8126001 and rs2618566) were in loci identified as liver tissue caQTL^30,58. Because base editing only alters a single variant in a set of GWAS candidates in linkage disequilibrium, this analysis supports at least partial causality to the tested variant, although in some cases bystander edit effects cannot be ruled out.

We performed deeper characterization of three variants for which editing alters both LDL-C uptake and chromatin accessibility. rs704 is a VTN missense variant and is the only variant in a credible set from the LDL-C GWAS⁵⁸. The other two variants are in gene promoters (rs35081008 is in the ZNF329 promoter, and rs8126001 is in the shared OPRL1 and RGS19 promoter (Fig. 4d)) and have moderate probability of causality from GWAS evidence (SuSiE fine-mapping PIP = 0.49 for rs35081008, PIP = 0.25 for rs8126001)⁵⁹. None of these most proximal target genes have been previously found to alter LDL-C uptake, nor do they show significance in LDL-C burden analyses³⁵.

We confirmed through quantitative PCR analysis with reverse transcription that editing the minor to major alleles of rs35081008 and rs8126001 leads to increased expression of ZNF329 and OPRL1, respectively (Fig. 4e,f), consistent with the increased chromatin accessibility induced by these edits. rs35081008 is heterozygous in HepG2 cells, and we used two linked ZNF329 intronic variants to assess allele-specific expression. In wild-type HepG2 cells, only 2% of ZNF329 transcripts derive from the minor allele haplotype (Fig. 4e), consistent with the status of rs35081008 as a liver eQTL⁴⁸. Editing rs35081008 from minor to major alleles increases expression of this haplotype to 35% of total transcripts (Fig. 4e), providing further evidence that rs35081008^Maj results in upregulated ZNF329 expression.

We then performed CRISPRa^46,47 and CRISPRi⁴⁸ to assess whether altered expression of the four candidate target genes alters LDL-C uptake (Methods). CRISPRa induction of VTN and ZNF329 significantly increased LDL-C uptake, and CRISPRi repression of VTN and OPRL1–RGS19 reduced LDL-C uptake (Fig. 4g). Our data are consistent with rs35081008^Min and rs8126001^Min decreasing ZNF329 and OPRL1 expression, respectively, each of which leads to decreased LDL-C uptake (Fig. 4h). We surmise that the VTN^T400M missense variant induced by the rs704^Min must have decreased VTN expression or function, consistent with prior evidence⁶⁰. For rs8126001, we adapted MotifRaptor⁶¹ to suggest that minor allele may impart stronger binding of zinc finger proteins ZNF333 and ZNF770 (Supplementary Note 6 and Supplementary Figs. 7 and 8), which could repress local chromatin^62,63. In summary, accurately quantifying variants’ impact on LDL-C uptake with BEAN, combined with experimental characterization, identifies genetic mechanisms underlying control of LDL-C levels.

BEAN quantifies LDLR tiling screen variant deleteriousness

We next adapted BEAN to the LDLR tiling library. Previous coding sequence base editing analyses have assumed single-allelic outcomes^45,49–51 (for example, all editable bases within a window are edited), which leads to erroneous amino acid mutation assignments, or have analyzed gRNA-level signal only with low-throughput validation of genotypes^13,15,21,22. We aimed to exploit the combination of dense tiling afforded by ABE8e–SpRY and reporter editing outcomes to model the effects of coding variants more accurately.

The LDLR tiling screen showed high coverage of edited nucleotides and amino acids (92% of targetable nucleotides and 74% of the 860 LDLR amino acids were edited at >10% frequency by at least one gRNA in the reporter; Supplementary Fig. 9). Each gRNA produced an average of 2.6 distinct alleles, and each variant was covered by 5.8 gRNA species on average (Supplementary Fig. 10). As opposed to the LDL-C GWAS analysis in which each gRNA was evaluated based on its editing frequency at a single target position, we adapted BEAN to incorporate multi-allelic outcomes, modeling the phenotype induced by a gRNA as the mixture of allelic phenotypes resulting from the gRNA’s activity, given the potential to observe multiple non-wild-type alleles per gRNA (Supplementary Note 2 and Fig. 5a). Moreover, base edits in coding regions are aggregated into amino acid-level variants. In addition to the output metrics described for the LDL-C GWAS screen, BEAN additionally outputs several per-variant metrics related to the confidence of BEAN scores (Supplementary Note 7 and Extended Data Fig. 6).

Fig. 5 |. Dissection of LDLR variant effects through BEAN modeling of a saturation tiled base editing screen.

a, BEAN model for coding sequence tiling screens. Coding region editing efficiencies are aggregated into amino acid-level mutations. Phenotypes of gRNA species with multi-allelic outcomes are modeled as the Gaussian mixture of allelic phenotypes (Methods). b, Scatterplot of estimated variant effect sizes and z scores, colored by ClinVar annotation. Point size corresponds to effective editing rate. Splice site variants are outlined in red. Selected deleterious variants without ClinVar P/LP annotations are labeled; labels are colored by protein domain. c, Ridge plot of BEAN z-score distributions of positive splicing control, synonymous and missense variants. d, Ridge plot of BEAN z-score distributions of ClinVar variants by clinical significance annotations. e, AUPRC of classifying ClinVar P/LP versus B/LB variants. Crosses show the metrics using all four replicates with no failing samples. Box plot shows the metrics of six two-replicate combinations of the replicates. f, LDLR protein domain structure adopted from Oommen et al.⁸⁷. EGF, epidermal growth factor. g, BEAN z scores for variants in the seven LDLR class A repeat domains aligned with the Pfam profile HMM logo by Skylign⁸⁸. Highly conserved cysteines are highlighted in gray. Ref AA, reference amino acid. h, Scatterplot of LDLR class A repeat missense variant BEAN z scores and ΔPfam profile HMM scores. Higher ΔPfam scores correspond to substitution from highly conserved to rarely observed amino acids (r = 0.569, ρ = 0.431). Alt, alternate; ref, reference. i, Comparison of mean statin-adjusted LDL-C levels and BEAN z scores for variants observed in the UKB and base editing (r = −0.454, ρ = −0.400). j, Comparison of mean statin-adjusted LDL-C levels and BEAN-FUSE scores for variants observed in the UKB and base editing (r = 0.512, ρ = 0.497). k, Predicted LDL-C levels of observed missense variants using BEAN-FUSE and PhyloP scores with tenfold cross-validation, compared with mean statin-adjusted LDL-C levels in the UKB (r = 0.608, ρ = 0.475). l, Box plots of BEAN-FUSE functional scores for UKB individuals with variants observed in our base editing screen with or without CAD. m, Box plots of statin-adjusted LDL-C levels of UKB individuals with variants observed in our base editing screen with or without CAD. The P value shown was determined by two-sided Wilcoxon rank-sum test. r, ρ, Pearson, Spearman correlation coefficient; RMSE, root-mean-squared error. For statistics in box plots, please see Statistical note for plots.

A total of 2,182 distinct variants were assessed, including 874 missense coding variants. BEAN assigned significant z scores (<−1.96, equivalent to 95% CI not covering 0) to 145 variants (Supplementary Table 9), 131 of which decreased LDL-C uptake (Fig. 5b). While 17 of 34 splice site variants significantly altered LDL-C uptake (Fig. 5c), 47 variants that significantly decreased LDL-C uptake are annotated in ClinVar³⁴ as pathogenic or likely pathogenic (P/LP), while 17 are ClinVar VUS or conflicting variants and none are ClinVar benign or likely benign (B/LB), indicating that BEAN can reliably estimate the pathogenicity for variants from base editing screens (Fig. 5d).

We compared the variant classification performance of BEAN to that of other available screen analysis methods^49–52 (Supplementary Note 5). As in the LDL-C GWAS screen, BEAN had better performance than any other method with a high AUPRC of 0.88 (Fig. 5e and Extended Data Fig. 7) and again outperformed BEAN-Reporter and BEAN-Uniform.

To gain insight into mechanisms by which these variants disrupt LDL-C uptake, we examined BEAN z scores for variants that reside within conserved functional domains (Fig. 5f,g and Extended Data Fig. 8). LDLR contains seven highly conserved LDLR class A repeats that bind to LDL, each of which is structurally anchored by six highly conserved cysteines that form three disulfide bonds⁶⁴. As expected, many of the missense variants with the strongest effects on LDLR function disrupt these cysteines (Fig. 5g and Supplementary Fig. 11). We further compared the BEAN z score of every installed variant with its change in amino acid conservation score from the Pfam profile HMM⁶⁵ (Methods) and observed moderate concordance (Spearman ρ = 0.43, Pearson r = 0.57; Fig. 5h).

We next asked whether BEAN scores were associated with statin-adjusted LDL-C levels⁶⁶ in the UKB⁶⁷ for individuals with paired exome sequencing and lipid level data. After quality control (Methods), there were 76 distinct LDLR missense variants observed in our base editing data present in UKB patients. We observed moderate concordance between the average patient LDL-C and BEAN scores for these variants (Spearman ρ = 0.40, Pearson r = 0.45; Fig. 5i), suggesting that BEAN scores are accurately and quantitatively associated with these human in vivo measurements.

As our base editing screen does not assay all possible mutation types per position, we used the Functional Substitution Estimation (FUSE)⁶⁷ pipeline to impute the impact of unobserved variants within a residue in which at least one variant is scored (BEAN + FUSE score; Methods). Applying FUSE to the 76 UKB variants with observed base editing data, BEAN-FUSE showed improved correlation with UKB patient LDL-C levels (Spearman ρ = 0.50, Pearson r = 0.51; Fig. 5j). BEAN-FUSE correlation with UKB patient LDL-C levels was robust but lower for all 358 missense LDLR variants with lipid measurements (Spearman ρ = 0.37, Pearson r = 0.35; Extended Data Fig. 9a). We further combined BEAN-FUSE scores and PhyloP 100way vertebrate conservation scores using XGBoost regression⁶⁸, achieving stronger correlation with UKB patient LDL-C levels than either BEAN-FUSE or PhyloP alone for the 76 variants observed in the base editing screen (Spearman ρ = 0.48, Pearson r = 0.61; Fig. 5k and Extended Data Fig. 9c) and for 358 variants with BEAN-FUSE scores (Spearman ρ = 0.37, Pearson r = 0.31; Extended Data Fig. 9b,d).

Individuals with pathogenic FH variants are at higher risk of CAD, even after controlling for LDL-C levels⁶⁹. We asked whether CAD incidence in patients with LDLR variants could be stratified by functional scores. We found that, for individuals with rare LDLR variants, functional scores processed by BEAN-FUSE were significantly higher for patients with prevalent or incident CAD (Wilcoxon rank-sum test, P = 0.0479; Fig. 5l) with more robust stratification than statin-adjusted LDL-C values for individuals with variants covered in the screen (Wilcoxon rank-sum test, P = 0.112; Fig. 5m and Supplementary Fig. 12). This demonstrates the advantage of quantifying genetic risk, which has a lifelong impact on LDL-C levels, over the snapshot provided by a single LDL-C measurement. Overall, we show that activity-normalized base editing screening can yield accurate quantitative estimation of LDLR variant pathogenicity in a large human cohort.

Structural basis of LDLR missense variants

We further analyzed 145 LDLR missense variants prioritized by BEAN (95% CI of z not covering 0) to gain insight into mechanisms of their pathogenicity. Variants that most severely abrogated LDL-C uptake included those that disrupt the start codon, disulfide bond-forming cysteine mutations in LDLR class A repeats and epidermal growth factor (EGF)-like domains, and mutations in the signal peptide. More in-depth characterization of variants that increase and decrease LDL-C uptake is provided in Supplementary Note 8.

We noticed that an appreciable number of deleterious variants that lack ClinVar pathogenic designation reside in the six LDLR class B repeats, which form a propeller-like structure involved in LDL release following its endocytosis. We computationally predicted^70–72 variant impact on structural stability (ΔΔG) and interatomic interactions (Methods and Supplementary Table 10). We found that the 26 significant LDLR class B variants induced more destabilizing effects, disrupted more hydrophobic interactions, had lower relative solvent accessibility (RSA)⁷³ (0.041 of the maximum RSA) and had higher wild-type residue depth than the other observed variants in this region (Fig. 6a–d and Supplementary Fig. 13). Collectively, these observations strongly indicate that these significant LDLR class B repeat variants are predominantly buried within the protein core where they engage in extensive hydrophobic interactions essential for protein folding. Moreover, we found a conserved interaction across repeat domains in which a tyrosine (aligned position 5 in Extended Data Fig. 8b) holds neighboring propeller blades together through interactions with a hydrophobic residue of the neighboring repeat (Fig. 6e,f). We identified five such variant pairs (Y442C with V481A, Y442C with V468A, Y489H with M531T, Y532C with L568P and Y576H with V618A), where all nine positions had at least one variant that weakened their hydrophobic interaction and had a significant BEAN z score. Among the top-ranked unannotated or ClinVar VUS and conflicting variants within LDLR class B repeats, the six most significant variants (L426P, Y489H, M531T, I566T, S584P and Y576H) all disrupt residues that hold the propeller blades together through hydrophobic interactions (Fig. 6g–i and Extended Data Fig. 10), while variants at these positions that conserve hydrophobicity show less severe impact (Fig. 6j). In summary, structural analysis of rare LDLR variants identified by BEAN highlights a central role for hydrophobic interactions in adhering the LDLR class B repeat domains.

Fig. 6 |. Deleterious variants in LDLR class B repeats weaken hydrophobic interactions.

a–d, Box plots of predicted ΔΔG (a), change in the number of hydrophobic interactions in the mutated structure (b), RSA (c) and residue depth (d) of 26 significant variants (z < −1.96) and the rest of the 259 variants (z > −1.96) observed in LDLR class B repeats. P values of two-sided Wilcoxon rank-sum test are denoted. Mut, mutated. e, Predicted protein structure of the LDLR class B repeat domain. Conserved interactions involving tyrosine of which mutation showed significant BEAN scores. Simplified interaction types and distance flags as annotated by Arpeggio are shown in the legend. f, BEAN z scores of positions with conserved hydrophobic residues are shown along with the LDLR class B repeat Pfam HMM logo. g–i, Local atomic interaction in wild-type and mutated structures for ClinVar conflicting variants or VUS L426P, M652T and Y532C. Residues in the variant positions are colored by the reference amino acids. Residues that interact with the variant position are shown. Variant positions and interacting residues are colored by elements (O, red; N, blue; S, yellow). j, Contour plot of BEAN z scores against predicted ΔΔG values for 872 missense variants. Positions with distinct observed missense variants that disrupt and conserve hydrophobic side chains are connected by dashed lines. For statistics in box plots, please see Statistical note for plots.

Discussion

In this work, we develop a framework to improve variant effect quantification in base editing screens by accounting for variable genotypic outcomes. The framework is straightforward to apply, simply requiring synthesis of gRNA–reporter pairs, which can be cloned and screened using standard experimental procedures. We note that, because this approach requires, for newly designed libraries, longer oligonucleotides than gRNA-only screening and requires paired-end sequencing, the cost is increased 10–50% for these steps, although the increment is minimal in context of the overall cost of screening. This framework should prove particularly useful when employing new CRISPR enzymes and deaminases, as it allows for simultaneous characterization of editing preferences and phenotypic screening. While the screens described here used flow cytometric phenotypic readouts, BEAN should also be suitable in dropout and enrichment paradigms by performing reporter analysis at an early time point. We provide an open-source implementation of BEAN in the comprehensive Python package bean (https://github.com/pinellolab/crispr-bean)⁷⁴ with end-to-end functionality.

We show that incorporation of reporter editing outcomes and accessibility in our model improves classification. Performance may be further improved by more thorough understanding of the effect of the epigenetic state on base editing, as has been achieved for Cas9 nuclease editing outcomes⁴³. It is also worth noting that BEAN currently does not consider off-target editing, which may affect the evaluation of phenotypic impacts for certain gRNA species.

We used BEAN to uncover common variants that modulate LDL-C levels by altering liver cell expression and/or function of three previously unappreciated genes, OPRL1, VTN and ZNF329. It is unclear why individuals with rare deleterious variants in these genes do not have altered LDL-C levels, but given that these genes have evidence of selective constraint⁷⁵, we speculate that variants that are tolerated in the population may have relatively weak functional and phenotypic effects⁷⁶.

We additionally demonstrate that dense activity-normalized base editing screens can improve characterization of coding variants in LDLR. Through the expanded coverage of ABE8e–SpRY and the sharing of information across adjacent gRNA species to boost power, we obtain accurate phenotypic measurements for an average of one variant per amino acid, considerably more than previous screens^{13,15–18,20–23,25–27}. Although this resolution is less than that of deep mutational scanning^4,77, base editing is considerably less work and cost intensive and far more scalable, allowing assessment of sets of genes in a single experiment. Notably, we show that BEAN scores for LDLR variants, most of which lack prior clinical annotation, are associated with quantitative serum LDL-C measurements in the UKB. Whereas functional assays are accepted as evidence of pathogenicity or benignity in clinical variant interpretation guidelines^78,79, these approaches have not been designed to offer risk predictions for quantitative traits. Given that CAD risk depends proportionally on the level of lifelong serum LDL-C exposure^80,81, the quantitative risk assessment we have shown for LDLR variants may have clinical utility by identifying variants that confer more moderate but significant effects. We additionally show, albeit in a small cohort, that LDLR functional scores are associated more closely with CAD incidence than LDL-C levels. This result is consistent with the residual CAD risk of patients with FH even after LDL-reductive therapy⁸² or controlling for LDL levels⁶⁹ and reinforces the value of LDLR pathogenicity analysis above and beyond LDL-C monitoring.

We show that base editing-derived functional scores correlate moderately with evolutionary conservation as well as structural disruption. We show preliminary evidence that integrating BEAN and PhyloP scores improves prediction of LDL-C levels. We anticipate that principled integration of distinct lines of evidence including data from multiplexed assays of variant effect will improve pathogenicity prediction^83–85, because each evidence type is an imperfect surrogate for the effects of a variant over the lifespan of an individual. In conclusion, activity-normalized base editor reporter screening with BEAN markedly improves variant impact quantification from CRISPR base editing screens, promising to accelerate the characterization of human disease-associated variants in their native genomic context.

Methods

Ethics statement

Human participant research was reviewed by the Mass General Brigham Institutional Review Board (protocol 2020P002093) and is under authorization from the UKB (project 41250).

Cloning base editor plasmids

To clone lentiviral plasmids used for base editor and gRNA–reporter expression, we subcloned the lentiCRISPR v2 plasmid (Addgene plasmid 52961; http://n2t.net/addgene:52961; RRID Addgene 52961)⁸⁹ to make the following changes using NEBuilder (New England Biolabs)-based cloning: we replaced the original gRNA hairpin with the FE hairpin⁹⁰. We swapped the Cas9 construct with a D10A nickase mutant of the Cas9 variant SpRY³⁸ using Addgene plasmid 139989 (http://n2t.net/addgene:139989; RRID Addgene 139989 (ref. 38)). We added sequences for either N-terminal ABE8e³⁷ from a synthesized sequence block or N-terminal AID and C-terminal UGI²⁹ using plasmid p2T-CMV-AIDmax-BlastR, a gift from D. Liu (Department of Chemistry and Chemical Biology, Harvard University) (Addgene plasmid 152994; http://n2t.net/addgene:152994; RRID Addgene 152994)²⁹.

To create HepG2 cells stably expressing base editors, we cloned editors into an existing pLenti plasmid backbone that allows for ORF fusion with P2A-BFP and hygromycin selection. We cloned pLenti_ ABE8e-SpRY-P2A-BFP_HygroR by transferring the sequence for the ABE8e–SpRY editor from lentiCRISPRv2FE-ABE8e-SpRY, and we cloned pLenti_AID-BE5-SpRY-P2A-BFP_HygroR by transferring the sequence for the AID-BE5 editor from lentiCRISPRv2FE-AIDBE5-SpRY.

The sequences of the engineered regions of these vectors are shown in Supplementary Data 1. We have submitted the ABE plasmids (lentiCRISPRv2FE-ABE8e-SpRY and pLenti_ABE8e-SpRY-P2A-BFP_HygroR) to Addgene because we obtained more robust data from ABE8e screening.

Establishing cell lines

HepG2 cells were obtained from the American Type Culture Collection and were authenticated by confirming that heterozygosity at dozens of loci and RNA-seq profiles match ENCODE profiles. For all experiments, HepG2 cells were kept in culture with DMEM medium with 10% FBS and were grown in an incubator at 37 °C with normal ATM pressure, 20% O₂ and 5% CO₂. HepG2 cells were infected with lentiviral constitutive base editor vectors pLenti_ABE8e-SpRY-P2A-BFP_HygroR and pLenti_AID-BE5-SpRY-P2A-BFP_HygroR. After selection with 66 μg ml⁻¹ hygromycin B, cells were sorted twice to enrich for BFP expression.

Base editing screening

The gRNA libraries were cloned into either CRISPRv2FE-ABE8e-SpRY-BsrGI or CRISPRv2FE-AIDBE5-SpRY-BsrGI. Libraries were amplified using NEBNext Ultra II Q5 Master Mix, cloned using NEBuilder HiFi DNA Assembly Master Mix and electroporated into Endura competent cells (Biosearch Technologies) for propagation. Lentivirus was produced in HEK293T cells (ATCC CRL-3216) for each library using TransIT-Lenti transfection reagent, titered and incubated with 6.25 × 10⁶ ABE8e–SpRY stable HepG2 cells and AID-BE5–SpRY stable HepG2 cells, respectively, per replicate at a multiplicity of infection of 0.3–0.5. Five or more biological replicate screens were performed for each library. After 24 h, medium containing lentivirus was removed, and fresh DMEM medium with 2 mM VPA was added to allow for construct integration and promote base editing. After another 48 h of treatment with medium and VPA, medium with 500 ng ml⁻¹ puromycin was added, and cells underwent selection for the next 5–7 d and were split as needed.

After complete selection as defined by complete death of a concurrent untransduced HepG2 control, cells were started on the 2-d LDL uptake protocol: on day 0, library cells were split, counted and replated onto 15-cm plates at 1 × 10⁵ cells per cm². On the evening of day 1, DMEM medium was removed and replaced with Opti-MEM to induce overnight serum starvation. On the morning of day 2, 2.5 μg ml⁻¹ BODIPY FL-LDL (Thermo Fisher Scientific) and Opti-MEM were added and incubated with the cells for 4–6 h. After this incubation period, cells in plates were trypsinized, stained with 50 ng ml⁻¹ DAPI and sorted based on BODIPY-LDL fluorescence levels into four bins (top 20%, top 20–40%, bottom 20% and bottom 20–40%). gDNA was then collected from each sorted population as well as from an unsorted bulk population and prepared for NGS.

Library preparation and next-generation sequencing

gDNA collected from each population was used as the input for a PCR1 reaction aimed at amplifying the integrated construct spanning the gRNA as well as adding different inline PE1 barcodes to specific samples for downstream analysis. The ideal total gDNA input per sample for PCR1 was 20 μg DNA, but if less than 20 μg gDNA was collected for a specific sorting population, then the total gDNA yield was input into PCR1. NEBNext Ultra II Q5 Master Mix was used. Please refer to Supplementary Table 11 for specific PCR1 primers used. Following PCR1, reactions were individually PCR purified using a standard QIAquick PCR Purification Kit. Next, a qPCR2 was performed from 0.25 μl of each purified sample in a 15-μl qPCR reaction to determine the optimum number of cycles for PCR2, and the products of qPCR2 were run on a gel to confirm lack of primer dimers. PCR2 cycle counts were chosen to be two to three cycles less than the qPCR2 Ct for the corresponding sample, with the minimum number of PCR2 cycles being seven. To set up PCR2, half of each sample’s purified PCR1 product was then used with NEBNext Ultra II Q5 Master Mix (please refer to Supplementary Table 11 for specific PCR2 primers used). After PCR2, samples were PCR purified again and were then run on a 2200 Agilent TapeStation to quantify the product as well as any unwanted byproducts and primer dimers that might have been generated during NGS preparation. Samples were then pooled based on their molarity, and the pool was purified to remove unwanted products using SPRIselect beads from Beckman Coulter. After bead purification, the pooled sample was sequenced using the Illumina NextSeq with paired-end sequencing and >50 nucleotides for read 1 and >36 nucleotide for read 2.

Comparison of endogenous and reporter editing

Endogenous sublibraries A–D targeting four distinct loci across LDLR (Supplementary Table 5) were cloned and tested identically as in the screens (see ‘Base editing screening’ section) but in six-well format. After gDNA isolation, a reporter-based library preparation for NGS was performed (Library preparation and next-generation sequencing), together with an additional library preparation for amplifying endogenous editing sites: PCR1 samples to determine endogenous editing were set up with Ultra II Q5 Master Mix and 2.5 μg gDNA input and a unique primer mix for each of the four LDLR endogenous libraries. These primer mixes contained three unique R1 forward primers and one common R2 reverse primer that would allow amplification of segments from the endogenous LDLR containing all desired editing sites targeted within that library. Upon completion of PCR1, samples were PCR purified using a QIAquick PCR Purification Kit and prepared following the standard library preparation protocol above. For specific primers used, please see Supplementary Table 11. After the preparation of these endogenous samples, both endogenous and reporter samples were run on a 2200 Agilent TapeStation and pooled and purified accordingly to prepare for NGS (Library preparation and next-generation sequencing).

Endogenous target site reads were mapped to the reference amplicon sequences using CRISPResso2 (ref. 91) (version 2.2.9) with base editing mode and a custom mismatch score matrix that tolerates A-to-G mutation generated by the ‘bean-count’ command of bean software⁷⁴ that implements all the computational functions of the proposed BEAN workflow as a Python package. Bean is available at https://github.com/pinellolab/crispr-bean/. The paired gRNA and reporter library is mapped to the expected gRNA and reporter sequences using the ‘bean-count’ function of the bean package. Position-wise reporter base edits are tested for significance against control data without editing using the function ‘bean.annotate.filter_alleles.filter_alleles’. This function conducts Fisher’s exact test and was used to filter for edits with Bonferroni-corrected P value < 0.05 and odds ratio > 5.

For comparison of the relative editing rates of multiple edits installed by a gRNA, the maximal editing rate among the installed edits by a gRNA was taken to divide the nucleotide-level editing rate of all variants installed by the gRNA. This editing rate normalized by the maximal editing rate was called ‘relative editing rate’. Variants with less than 5% maximal editing rate were excluded for the stability in calculating relative editing rates.

Base editing screen analysis with BEAN

We developed BEAN, a computational pipeline that handles gRNA and editing outcome quantification, quality control, editing preference profiling and variant impact quantification. For variant impact quantification, BEAN models the screen procedure using Bayesian networks. A detailed description of the BEAN pipeline is provided in Supplementary Note 2.

Prediction of editing outcomes with BE-Hive

We used the Python implementation of BE-Hive²⁹ (https://github.com/maxwshen/be_predict_bystander/commit/31aadd) to predict editing outcomes of the LDLR tiling library. We initialized the model with ‘mES’ as cell type and ‘ABE8’ as the editor due to the lack of a HepG2 cell type model. For each spacer, we extracted a 50-nucleotide-long sequence around the starting position of the spacer in the hg38 genome, with 20 nucleotides before the spacer starts and 30 nucleotides after, on the same strand. These 50-nucleotide sequences were used as input to BE-Hive to predict likely editing outcomes. To calculate allele-level edit rates for each spacer, we summed up the probability of any editing outcomes with the same editing patterns in positions 0–18 (0-based) relative to the start of the spacer. Similarly, to calculate base-level edit rate, we summed up the probability of any editing outcomes with identical base edits in positions 3–8, relative to the start of the spacer.

Quantification of editing rates

We quantified per-gRNA reporter and protospacer self-editing rates as follows. For the variant screen where we knew the designated target position, the number of reads with the position edited was divided by the number of total reads. For the tiling screen without set target position, mean editing rates of all editable bases within protospacer positions 3–8 (given 1 is the start position of the protospacer) were used.

Cloning and testing of individual gRNA species

We performed fluorescent LDL-C uptake profiling of each edited cell line mixed with an in-well control cell line in six biological replicates, allowing us to compare changes in LDL-C uptake with matched data from the screen. Oligonucleotides including protospacer sequences were ordered in the following format: GGAAAGGACGAAACACCG[19–20-bp protospacer: remove initial G for any 20-bp protospacer with one natively]GTTTAAGAGCTATGCTGGAAAC (Supplementary Table 11). Using NEBuilder HiFi DNA Assembly, ABE8e–Cas9NG-designated oligonucleotides were cloned into CRISPRv2FE-ABE8e-Cas9NG, while ABE8e–SPRY-designated oligonucleotides were cloned into CRISPRv2FE-ABE8e-SpRY-BsrGI. To make base edited cell lines, the gRNA constructs were packaged into lentivirus and transduced into HepG2 ABE8e–SPRY-BFP cells seeded at 4 × 10⁴ cells per cm² in six-well plates in two replicates with 8 μg ml⁻¹ Polybrene. Two days after transduction, cells were treated with 500 ng ml⁻¹ puromycin and selected for approximately 1 week. HepG2 base edited cells were seeded 1:1 with HepG2-mCherry cells to achieve a total density of 1.08 × 10⁵ cells per cm² in a 96-well plate in at least two technical replicates of two biological replicates and incubated overnight. The next day, the medium was replaced with Opti-MEM, and cells were incubated overnight. Approximately 4–6 h before flow cytometric analysis, cells were treated with 2.5 μg ml⁻¹ BODIPY-LDL in Opti-MEM. Cells were trypsinized and analyzed for the presence of mCherry and LDL uptake using a Beckman CytoFLEX flow cytometer, and data were analyzed with CytExpert version 2.5 and Cytobank version 10.3. LDL uptake of each base edited cell line was normalized to the LDL uptake of the mCherry cells within the same well. Differential LDL uptake between base edited and control cells was further normalized using data from the ABE8e and SPRY sgCTRL lines.

CRISPRi

Oligonucleotides including protospacer sequences (Supplementary Table 11) were ordered in the following format: GGAAAGGACGAAACACCG[19–20-bp protospacer: remove initial G for any 20-bp protospacer with one natively]GTTTAAGAGCTATGCTG-GAAAC were cloned into a pHR-U6-gRNAFE-Zim3-dCas9-P2A-Hygro backbone by NEBuilder HiFi DNA Assembly. To make CRISPRi cell lines, the gRNA constructs were packaged into lentivirus and transduced into HepG2 cells seeded at 4 × 10⁴ cells per cm² in 48-well plates in two replicates with 8 μg ml^–1 Polybrene. Two days after transduction, cells were treated with 125 μg ml⁻¹ hygromycin B and selected for approximately 1 week. LDL uptake experiments were performed as described above, seeding CRISPRi cell lines 1:1 with HepG2-tTA-BFP cells as the internal control.

CRISPRa

Oligonucleotides including protospacer sequences (Supplementary Table 11) were ordered in the following format: GGAAAGGACGAAACACCG[19–20-bp protospacer: remove initial G for any 20-bp protospacer with one natively]GTTTAAGAGCTAG-GCCAACATG. Using NEBuilder HiFi DNA Assembly, oligonucleotides were cloned into a pLenti U6–2xMS2gRNA MCPp65 PuroR backbone. To make CRISPRa cell lines, the gRNA constructs were packaged into lentivirus and transduced into HepG2 dCas9–10xGcn4-mChe + scFv-Sbno1-Nfe2l1-Krt40-BFP cells and seeded at 4 × 10⁴ cells per cm² in six-well plates in two replicates with 8 μg ml⁻¹ Polybrene. Two days after transduction, cells were treated with 500 ng ml⁻¹ puromycin and selected for approximately 1 week. LDL uptake experiments were performed as described above, seeding CRISPRa cell lines 1:1 with HepG2 wild-type cells as the internal control.

Pooled ATAC-seq

Lentiviral delivery of a pool of 20 ABE8e–SpRY gRNA species to HepG2 cells at a high multiplicity of infection was followed by ATAC-seq and paired gDNA collection in three biological replicates in standard and serum-starved conditions. We performed multiplexed PCR enrichment of the regions surrounding each of the 20 edited variants followed by targeted amplicon sequencing by NGS. Differential representation of an alternate allele in ATAC-seq relative to gDNA sequencing implies differential accessibility of the alternate allele as compared to the reference.

A pool of 20 gRNA species was cloned into CRISPRv2FE-ABE8e-Cas9NG or CRISPRv2FE-ABE8e-SpRY-BsrGI (Cloning and testing of individual gRNA species), packaged into lentivirus and transduced into HepGw2 ABE8e-SpRY-BFP cells seeded at 4 × 10⁴ cells per cm² in six-well plates in three replicates with 8 μg ml⁻¹ Polybrene. Cells were treated with VPA and selected with puromycin as in screens. Once selected, two sets of 1 × 10⁶ cells for each of the three replicates as well as an unedited control replicate were seeded in six-well plates. The next day, one well per replicate was fed DMEM with FBS and the other Opti-MEM (serum starved). Twenty-four hours later, cells in the wells were trypsinized, and 1 × 10⁵ cells were used for ATAC-seq, while the remaining cells were used for bulk gDNA isolation using the Purelink Genomic DNA Mini Kit (Life Technologies). ATAC-seq was performed using the Active Motif ATAC-Seq Kit according to the manufacturer’s instructions.

To obtain valid primers to amplify the loci surrounding the 20 target variants, Primer3 (ref. 92) was used to generate five candidate primer sets within ±150 nucleotides from each variant. http://primer-dimer.com was used to calculate a ΔG interaction matrix for all candidate primers. Primers with an average ΔG ≤ −7 were removed. Next, recursive pairwise filtering was performed to iteratively remove the primer with the worst ΔG interaction until no pairwise ΔG ≤ −7 remained. This recursive filtering was performed 300 times, and the run with the most primers remaining was used. The primer set for each variant with the highest minimum ΔG was selected. Primers were all ordered from IDT preceded by NNN to randomize initial nucleotides in NGS. We provide the amplicon sequence of 20 loci in Supplementary Table 11.

gDNA and ATAC-seq products from a total of eight samples (three experimental samples and one control sample, both in serum and starved conditions) were amplified using two primer pools, each composed of ten primer sets to synchronize the annealing temperature. Two and a half micrograms of gDNA or half of the ATAC-seq product was used in 100-μl reactions for 32 cycles (gDNA) or 35 cycles (ATAC-seq). The TapeStation was used to pool the two PCR products for each sample, and these 16 pools were used as input to the NEB-Next Ultra II DNA Library Prep Kit to prepare NGS libraries. Libraries were sequenced using 150-nucleotide single-end sequencing on the Illumina NextSeq.

Pooled ATAC-seq analysis

For each sample, ATAC-seq reads were mapped using Bowtie 2 (ref. 93) (version 2.5.1) with default options to each of the expected amplicon sequences (Supplementary Table 11) from the 20 loci that were amplified by PCR during library preparation. ‘bowtie2-build’ was used to build indices for the amplicon sequences for each of 20 loci, and reads were mapped onto the indices with default parameters. An in-house Perl script was used to parse the SAM output from Bowtie 2 and to demultiplex the reads by the locus they mapped to with default options. Demultiplexed reads were then profiled for the target base editing rate using CRISPResso2 (ref. 91) (version 2.2.9) and an average read quality cutoff of a Phred score of 30 and assigned base ‘N’ if per-base quality was lower than a Phred score of 20. For each variant, reads are assigned to the reference allele or the alternate allele based on the base identity at the target SNP position. In case there exists a neighboring variant that allows phasing, as the HepG2 line is heterozygous for the variant, the reads were counted per phase based on the identity of the neighboring variant. We note that, for two of the variants examined (rs3767844 and rs4390169), whether the base was the result of editing or was the reference allele was ambiguous (that is, variants were heterozygous in HepG2 cells and two reference alleles of A and G; therefore, we cannot assign reads with G in the variant position to the edited reference A or the unedited reference G). For the variants, we simply compared two observed bases and treated the effect as caQTL. For rs771555783, rs76895963 and rs116734477, edited reads were not detected due to insufficient representation of the loci and were thus excluded from the enrichment analysis.

We first identified the variants with significant editing observed in treatment samples compared to the control samples where base editors are not treated. This was carried out by assessing the significance of coefficient for is_treatment in the following binomial regression with the ‘GLM’ module of the Python statsmodels package⁹⁴ (version 0.12.1), where ${\text{Edited}}_j$ and ${\text{Undited}}_j$ are the read counts of edited and unedited variants in $\text{sample}j$, and ${\text{is\_treatment}}_j$ is the indicator variable for the $\text{sample}j$ being the treatment sample.

$$\begin{gathered} {\text{Edited}}_j\approx \text{Binomial}\left(p_j,{\text{Edited}}_j+{\text{Unedited}}_j\right) \\ \text{logit}\left(p_j\right)\approx 1+{\text{is\_treatment}}_j \end{gathered}$$

Significantly edited variants should show higher proportions of edited reads $\left(p_j\right)$ in treatment samples than in the control samples. For all significance testing, a Benjamini–Hochberg FWER value of 0.1 was used as the threshold, and multiple-testing correction was performed with the ‘stats.multitest.multipletests’ function of the Python statsmodels package⁹⁴ (version 0.12.1).

For the variants with significant observed editing, we calculated enrichment of editing in ATAC-seq compared to the gDNA sample, which indicates that editing opened the chromatin at the variant loci and increased its capture rate for ATAC-seq. The enrichment of the edited allele was calculated as the binomial regression coefficient of edited and unedited read counts for each variant. The proportion of the edited read count was regressed on whether the sequencing sample $j$ is from ATAC-seq $\left({\text{is\_ATAC}}_j=1\right)$ or gDNA $\left({\text{is\_ATAC}}_j=0\right)$, and the regression coefficient of ${\text{is\_ATAC}}_j$ was used as the accessibility enrichment of the variant editing. We conditioned for replicate- and condition-specific effects, along with the interaction effect between condition and ATAC-seq sample to examine whether the variant only alters accessibility under either one of two conditions (serum fed and serum starved, condition = 1 for serum fed).

$$\begin{gathered} {\text{Edited}}_j\approx \text{Binomial}\left(p_j,{\text{Edited}}_j+{\text{Unedited}}_j\right) \\ \text{logit}\left(p_j\right)\approx 1+{\text{is\_ATAC}}_j+{\text{condition}}_j\times {\text{is\_ATAC}}_j+{\text{condition}}_j+{\text{replicate}}_j \end{gathered}$$

We also calculated the caQTL effect of the variants that were heterozygous in HepG2 cells. Here, whether one allele had higher enrichment in an ATAC-seq sample was examined as the regression coefficient for ${\text{is\_ATAC}}_j$ of the following regression, again conditioned on experimental condition and replicate.

$$\begin{array}{c}{\text{Allele0}}_j\approx \text{Binomial}\left(q_j,{\text{Allele0}}_j+{\text{Allele1}}_j\right)\\ \text{logit}\left(q_j\right)\approx 1+{\text{is\_ATAC}}_j+{\text{condition}}_j\times {\text{is\_ATAC}}_j+{\text{condition}}_j+{\text{replicate}}_j\end{array}$$

Here, $\text{Allele}{0}_j$ and $\text{Allele}{1}_j$ are the read counts of alleles 0 and 1. The regression coefficient for ${\text{is\_ATAC}}_j$ is used as the accessibility enrichment of allele 0. When the enrichment is shown uniformly in major to minor alleles, the signs of enrichment values and confidence intervals calculated for the opposite direction are inverted.

Pfam profile HMM scores

Pfam profile HMM files of PF00057, PF00058 and PF00008 for LDLR class A repeat, LDLR class B repeat and EGF-like domain, respectively, were downloaded from Pfam⁶⁵ to generate sequence logo through Skylign⁸⁸ (version from August 2023), where the height of each position shows its information content and letter height shows the total height scaled by relative frequencies of the letters in the position. The match emission score from the profile HMMs was used to calculate the ΔPfam score. The match emission score is the negative log probability to observe the amino acid from multiple sequence alignment for a given position; thus lower score corresponds to high conservation and lower ΔPfam(reference − alternate) = −(Pfam_alternate − Pfam_reference) corresponds to higher reference amino acid conservation and lower chance to observe the alternate amino acid.

LDLR repeat domain alignment

LDLR class A repeats were aligned as shown in a previous study to align for all cysteine residues. Alignments for LDLR class B repeats and the EGF-like domain were obtained with Clustal Omega^95–97 by aligning domain sequences with seed alignments from Pfam PF00058 and PF00008.

UK Biobank data processing

Study participants.

The UKB⁹⁸ is a prospective cohort of over 500,000 individuals recruited between 2006 and 2010 of ages 40–69. Drawing from 469,803 participants with whole-exome sequencing data, we included 443,353 participants with available LDL-C measurements in this study. Patients with homozygous variants and participants with more than one rare variant across LDLR and with any rare variant in APOB and PCSK9 were not considered for these analyses to control for their potential contribution to the serum LDL-C level, retaining 9,819 individuals harboring 358 distinct LDLR missense variants.

Variant inclusion and quality control.

Exon coordinates were determined for LDLR, APOB and PCSK9 using MANE transcripts⁹⁹, with an additional five nucleotides retained upstream and downstream of each coding region to capture splice site variants. Exome sequencing was performed for UKB participants as previously described. Analysis was conducted on the Research Analysis Platform (https://ukbiobank.dnanexus.com). We extracted gene-level VCF files from the whole-exome sequencing joint-called pVCFs using BCFtools¹⁰⁰ (version 1.15.1) and the Swiss Army Knife app (version 4.9.1) and then normalized to flatten multi-allelic sites and align variants to the GRCh38 reference genome.

Variants in low-complexity regions, segmental duplications or other regions known to be challenging for NGS alignment or calling were removed from analysis (National Institute of Standards and Technology Genome in a Bottle Consortium¹⁰¹ difficult regions), as were variants with an alternate allele frequency greater than 0.1% in the UKB cohort. Further filtering removed variants for which more than 10% of samples were missing genotype calls and variants that did not appear in the UKB cohort. To mitigate differences in sequencing coverage between individuals who were sampled at different phases of the UKB project, variants were only retained in the final set if at least 90% of their called genotypes had a read depth of at least 10.

The canonical functional consequence of each variant was calculated using Variant Effect Predictor (version 99)¹⁰². Noncoding variants outside of essential splice sites were not considered in the analysis. Computational scores were provided by Variant Effect Predictor, including the PhastCons conservation score¹⁰³ (PhastCons100way_vertebrate). When multiple PhastCons conservation scores were available for a coding variant, the mean of the available scores was used.

Clinical endpoints and endophenotypic data.

Data on CAD and myocardial infarction were aggregated from hospital records (primary or secondary diagnosis), death registries (primary or secondary cause of death) and self-reported data. Age of onset was estimated based on date of onset and birth date when not directly provided, and individuals with uncertain or unavailable onset data were excluded.

Patient-level LDL-C values were ascertained from UKB data files. Estimated untreated LDL-C levels obtained using adjustments for lipid-lowering therapies were used in analyses, as described in the Supplementary Information.

ClinVar assertions.

ClinVar clinical assessments were identified from the tab-delimited version of ClinVar¹⁰⁴ released on 4 April 2023. In this analysis, we use ‘pathogenic’, ‘likely pathogenic’ and ‘pathogenic/likely pathogenic’ classifications as ‘P/LP’ collectively and ‘benign’, ‘likely benign’ and ‘benign/likely benign’ classifications as B/LB.

BEAN-FUSE scores

We made use of the FUSE pipeline⁶⁷ (version from 17 August 2023) to improve the estimation of variant functional effects and to impute effects of variants that had not been screened. FUSE makes use of related measurements within and across experimental assays, namely an amino acid substitution matrix derived from 24 deep mutational scanning datasets, to jointly estimate variant impacts.

After functional scores had been estimated by the BEAN pipeline, the full set of scores was processed by FUSE, which first collectively estimates the mean functional effect per amino acid residue position within the assay, using shrinkage estimation. FUSE then makes estimates for individual allelic variants within the amino acid residue position, based on a functional substitution matrix derived from deep mutational scanning data across many genes. The result is a full set of estimated variant functional effects for both (1) the original variants screened in the assay and (2) other possible variants that were not screened but fell within amino acid residues that had variants covered in the screen.

Prediction of UKB LDL-C level

The UKB LDL-C level of variants observed in the base editing and with high confidence $\left({\sigma }_{\mu }<0.5\right)$ was predicted using XGBoost⁶⁸ Python package version 1.7.5 and with the default option with tenfold cross-validation implemented in scikit-learn¹⁰⁵ version 1.3.0 ‘model_selection.cross_val_predict’. The LDL-C levels of UKB variants that were unobserved or not observed with enough confidence ${\sigma }_{\mu }>0.5$ were predicted by the XGBoost model that was trained on variants observed with ${\sigma }_{\mu }<0.5$.

Structural analysis

Protein structures were visualized using PyMOL¹⁰⁶ (version 2.5.2). RSA⁷³ and residue depth were calculated using the DSSP module in Biopython (version 1.79)¹⁰⁷ to capture the local 3D accessibility of residues. Wild-type atomic interactions between residues were calculated with Arpeggio⁷¹ (initial release) using the LDLR AlphaFold2 structure (positions 1–860) from the AlphaFold Protein Structure Database⁷². Additionally, interactions with calcium ions and saccharides were calculated using PDB structure 1N7D (positions 65–714 after renumbering according to UniProt⁹⁷ P01130). These interactions were also computed for mutant structures generated using MODELLER¹⁰⁸ 10.3. Subsequently, the change in interactions was determined by subtracting the interactions in the mutant from those in the wild type. Within each of the LDLR class B, LDLR class A and EGF-like domains, two-sided Wilcoxon rank-sum tests were conducted to compare the features calculated for deleterious variants (identified by BEAN z scores below −1.96) against those of other variants. DDMut⁷⁰ (initial release) is a deep learning model that predicts protein stability change induced by mutation, ΔΔG, based on the local atomic environment and interactions in wild-type and mutated residues. The LDLR AlphaFold2 structure was used as the input to predict the ΔΔG of variants observed in our LDLR tiling screen.

Molecular interactions were visually represented using a color-coded scheme to differentiate between interaction types. Hydrophobic interactions are depicted in ‘forest’; polar interactions are depicted in ‘orange’; carbonyl interactions are depicted in ‘blue’; hydrogen bonds are depicted in ‘red’; aromatic ring interactions including methionine sulfur–π, donor–π, cation–π and amide–ring interactions are depicted in ‘pale green’; undefined interactions are depicted in ‘cyan’; coordinate covalent bonds are depicted in ‘purple’; and ionic interactions are depicted in ‘yellow’. Moreover, line type specifies distance flag from Arpeggio output, and thin dashed lines represent van der Waals clashes, where the van der Waals radii between two atoms cause steric clashes. When such clashes co-occur with other interactions mentioned earlier, they are portrayed with dashed lines, using the color code corresponding to the additional interaction except for the undefined interaction type. Additionally, all ionic interactions and coordinate covalent bonds with ions are consistently represented by yellow and purple dashed thick lines. Other interactions are all represented by solid lines.

Statistical note for plots

For all box plots, the bounds of the boxes are interquartile ranges and the centers show medians. Whiskers of the box plot show 1.5-fold of interquartile ranges as default in seaborn.boxplot¹⁰⁹.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Extended Data

Extended Data Fig. 1 |. Base editor editing preference profile and context specificity.

Deamination motif and PAM-dependent editing preference of AID-BE5-SpRY from 7294 gRNAs and AID-BE5-Cas9NG from 7299 gRNAs with more than 9 read counts across any replicates of bulk samples. a) Context specificity of AID-BE5-SpRY are represented as sequence logos. The height of each base represents the relative editing efficiency with each base. b) Mean editing efficiency of AID-BE5-SpRY by protospacer position and PAM sequence. c) Context specificity of AID-BE5-CasNG is represented as sequence logos. The height of each base represents the relative editing efficiency with each base. d) Mean editing efficiency of AID-BE5-CasNG by protospacer position and PAM sequence.

Extended Data Fig. 2 |. Nucleotide-level editing comparison of reporter and endogenous locus.

a) Scatterplots comparing of per-nucleotide-level editing efficiencies between the reporter and endogenous target sites. All edits introduced by each of 49 gRNAs across four loci across 3 experimental replicates were plotted. Points are colored by the identity of nucleotide edit and gRNA. b) The same plot colored by gRNA strand. R; Pearson correlation coefficient. n; number of plotted editing rates.

Extended Data Fig. 3 |. BEAN plate diagrams.

Plate diagrams of a) BEAN b) BEAN-Reporter, c) BEAN-Uniform. X^b and all parameters with superscript b is not used for benchmark analyses.

Extended Data Fig. 4 |. LDL-C GWAS library classification task benchmark.

a) AUPRC plot for classifying positive splicing control variants against negative control variants. Metrics for all 5 replicates are shown as markers and metrics of 15 two-replicate subsamples among the 5 replicates are shown as box plots. Boxplot was plotted as described in the statistical note of the Methods section. b) Precision-Recall curve for classifying all positive control splice sites of against negative controls for all replicates with no failing samples. c) Precision-Recall curve for classifying splice sites of LDLR/MYLIP against negative controls for all replicates with no failing samples. d) Precision-Recall curve for classifying all positive control splice sites of against negative controls for 2-replicate subsample of the data. Mean Precision value for a recall across 15 subsample runs are plotted as solid line. e) Precision-Recall curve for classifying splice sites of LDLR/MYLIP against negative controls for 2-replicate subsample of the data. Mean Precision value for a recall across 15 subsample runs are plotted as solid line.

Extended Data Fig. 5 |. Comparison of inferred effect sizes of individually transfected LDL-C GWAS library gRNAs.

Scatterplot and Pearson correlation coefficients (R) of effect size estimates and the log fold change (LFC) of fluorescence signal following individual transfection of 22 gRNAs. R; Spearman correlation coefficient.

Extended Data Fig. 6 |. BEAN accurately estimates variant effect confidence from per-variant evidence in input data.

a-c) Scatterplot of 2,182 LDLR tiling library variants comparing a) n_norm and effective edit rates, b) effective edit rates and BEAN σ_μ, and c) n_norm and BEAN σ_μ. d) Histogram of effective edit rates of 76 LDLR tiling library variants with UKB LDL-C levels. Quartile bin cutoffs used to categorize variants are shown as dotted lines. e) Scatterplots of BEAN z-scores and statin-adjusted UKB LDL-C measurements for variants in each effective edit rate quartile bin. r and rho shows the Pearson and Sparman correlation coefficients, respectively.

Extended Data Fig. 7 |. LDLR tiling library classification task benchmark.

a) AUPRC of classifying Pathogenic/Likely Pathogenic from Benign/Likely Benign variants when using 4 replicates without failing samples and 6 2-replicates combinations among the replicates. Bounds and the center of the boxes are the interquartile ranges. Boxplot was plotted as described in the statistical note of the Methods section. b-e) Precision-recall curve of classifying b, d) Pathogenic/Likely Pathogenic c, e) Pathogenic from Benign/Likely Benign variants. Top panels (b, c) show classification when used 4 replicates without failing samples. Bottom panels (d, e) show when used 6 2-replicates combinations among 4 replicates without failing samples.

Extended Data Fig. 8 |. Comparison of functional impact and conservation within conserved LDLR domains.

Repeat domain alignments shown with BEAN z-score for a) LDLR class A repeat domain, b) LDLR class B repeat domain, c) EGF-like domains aligned with the Pfam profile HMM logo by Skylign, where the height of each position show its information content and letter heights show the total height scaled by relative frequencies of the letters in the position. For a), conserved cysteine residue position is highlighted and for b-c), consensus positions from Clustal Omega alignment output are highlighted in grey.

Extended Data Fig. 9 |. Expanded LDLR missense variant pathogenicity estimates with FUSE.

a) Scatterplot of all considered UKB variant mean statin-adjusted LDL-C level against imputed BEAN-FUSE score. b) Prediction outcome of unobserved variants with XGBoost model trained on observed UKB variants and mean statin-adjusted LDL levels. c-d) Correlation coefficients and root mean squared error (RMSE) for predicted and true UKB mean statin-adjusted LDL-C level for XGBoost model with FUSE score, PhastCons PhyloP conservation score, and both as the input in predicting LDL-C levels. c) Boxplot of metrics for prediction of observed variants with 10-fold cross validation (n = 10) d) Barplot of metrics for prediction of unobserved variants with model trained on observed variants (n = 1). r, ρ; Pearson, Spearman correlation coefficient, RMSE; Root mean squared error.

Extended Data Fig. 10 |. Local atomic interaction in wild type and mutated structure for selected variants in LDLR class B repeat domain.

a–k, Residues with interaction with the variant position are shown. Variant positions and interacting residues are colored by the reference amino acid and atomic elements (O: red, N: blue, S: yellow). Ref AA; reference amino acid.

Data availability

The processed data used in this study have been deposited at Zenodo (https://doi.org/10.5281/zenodo.10139794), and primary sequencing data are available at the Sequence Read Archive under accession PRJNA1042659. Controlled access, patient-level data from the UKB may be requested at https://ams.ukbiobank.ac.uk/ams/. Source data are provided with this paper.